arXiv subjects

Xiawei Wang

Publications and source records attributed to Xiawei Wang.

15 recordsLinked to original sources

arXiv2026

ASSERT: A Measurement Pipeline for GenAI Audits

Audits of generative AI (GenAI) systems often summarize behavior as a reported rate: how often the audited system complies with policy. Researchers and stakeholders use that rate to compare systems, track regressions, and gate deployment. A reported rate reflects both the system under audit and the measurement choices behind it, so a change in the rate can leave it unclear whether the system or those choices moved. We introduce ASSERT, a specification-driven measurement pipeline for GenAI audits that ties each reported rate to a written specification of the measurement choices used to produce it. ASSERT helps draft a behavioral rubric and test cases, then runs the audit against a GenAI system and returns a reported rate. In a case study on conversational deception, we observe that the reported rate moves substantially with the dialogue setup, the simulated user, the judge, and the evidence bar for non-compliance. These measurement choices substantially change the reported rate and can reorder GenAI system rankings. Because each reported rate is tied to an explicit specification, differences across audits are easier to attribute and interpret.

cs.CL↗

arXiv2026

GraphRAG-IRL: Personalized Recommendation with Graph-Grounded Inverse Reinforcement Learning and LLM Re-ranking

Personalized recommendation requires models that capture sequential user preferences while remaining robust to sparse feedback and semantic ambiguity. Recent work has explored large language models (LLMs) as recommenders and re-rankers, but pure prompt-based ranking often suffers from poor calibration, sensitivity to candidate ordering, and popularity bias. These limitations make LLMs useful semantic reasoners, but unreliable as standalone ranking engines. We present \textbf{GraphRAG-IRL}, a hybrid recommendation framework that combines graph-grounded feature construction, inverse reinforcement learning (IRL), and persona-guided LLM re-ranking. Our method constructs a heterogeneous knowledge graph over items, categories, and concepts, retrieves both individual and community preference context, and uses these signals to train a Maximum Entropy IRL model for calibrated pre-ranking. An LLM is then applied only to a short candidate list, where persona-guided prompts provide complementary semantic judgments that are fused with IRL rankings. Experiments show that GraphRAG-IRL is a strong standalone recommender: IRL-MLP with GraphRAG improves NDCG@10 by 15.7\% on MovieLens and 16.6\% on KuaiRand over supervised baselines. The results also show that IRL and GraphRAG are superadditive, with the combined gain exceeding the sum of their individual improvements. Persona-guided LLM fusion further improves ranking quality, yielding up to 16.8\% NDCG@10 improvement over the IRL-only baseline on MovieLens ml-1m, while score fusion on KuaiRand provides consistent gains of 4--6\% across LLM providers.

cs.IR↗

arXiv2025

Jailbreak Distillation: Renewable Safety Benchmarking

Large language models (LLMs) are rapidly deployed in critical applications, raising urgent needs for robust safety benchmarking. We propose Jailbreak Distillation (JBDistill), a novel benchmark construction framework that "distills" jailbreak attacks into high-quality and easily-updatable safety benchmarks. JBDistill utilizes a small set of development models and existing jailbreak attack algorithms to create a candidate prompt pool, then employs prompt selection algorithms to identify an effective subset of prompts as safety benchmarks. JBDistill addresses challenges in existing safety evaluation: the use of consistent evaluation prompts across models ensures fair comparisons and reproducibility. It requires minimal human effort to rerun the JBDistill pipeline and produce updated benchmarks, alleviating concerns on saturation and contamination. Extensive experiments demonstrate our benchmarks generalize robustly to 13 diverse evaluation models held out from benchmark construction, including proprietary, specialized, and newer-generation LLMs, significantly outperforming existing safety benchmarks in effectiveness while maintaining high separability and diversity. Our framework thus provides an effective, sustainable, and adaptable solution for streamlining safety evaluation.

cs.CL↗

arXiv2025

DeCafNet: Delegate and Conquer for Efficient Temporal Grounding in Long Videos

Long Video Temporal Grounding (LVTG) aims at identifying specific moments within lengthy videos based on user-provided text queries for effective content retrieval. The approach taken by existing methods of dividing video into clips and processing each clip via a full-scale expert encoder is challenging to scale due to prohibitive computational costs of processing a large number of clips in long videos. To address this issue, we introduce DeCafNet, an approach employing ``delegate-and-conquer'' strategy to achieve computation efficiency without sacrificing grounding performance. DeCafNet introduces a sidekick encoder that performs dense feature extraction over all video clips in a resource-efficient manner, while generating a saliency map to identify the most relevant clips for full processing by the expert encoder. To effectively leverage features from sidekick and expert encoders that exist at different temporal resolutions, we introduce DeCaf-Grounder, which unifies and refines them via query-aware temporal aggregation and multi-scale temporal refinement for accurate grounding. Experiments on two LTVG benchmark datasets demonstrate that DeCafNet reduces computation by up to 47\% while still outperforming existing methods, establishing a new state-of-the-art for LTVG in terms of both efficiency and performance. Our code is available at https://github.com/ZijiaLewisLu/CVPR2025-DeCafNet.

cs.CV↗

arXiv2024

Improving Lung Cancer Diagnosis and Survival Prediction with Deep Learning and CT Imaging

Lung cancer is a major cause of cancer-related deaths, and early diagnosis and treatment are crucial for improving patients' survival outcomes. In this paper, we propose to employ convolutional neural networks to model the non-linear relationship between the risk of lung cancer and the lungs' morphology revealed in the CT images. We apply a mini-batched loss that extends the Cox proportional hazards model to handle the non-convexity induced by neural networks, which also enables the training of large data sets. Additionally, we propose to combine mini-batched loss and binary cross-entropy to predict both lung cancer occurrence and the risk of mortality. Simulation results demonstrate the effectiveness of both the mini-batched loss with and without the censoring mechanism, as well as its combination with binary cross-entropy. We evaluate our approach on the National Lung Screening Trial data set with several 3D convolutional neural network architectures, achieving high AUC and C-index scores for lung cancer classification and survival prediction. These results, obtained from simulations and real data experiments, highlight the potential of our approach to improving the diagnosis and treatment of lung cancer.

eess.IV↗

arXiv2019

Non-thermal emission from the interaction of magnetized exoplanets with the wind of their host star

We study the non-thermal emission from the interaction between magnetized Jupiter-like exoplanets and the wind from their host star. The supersonic motion of planets through the wind forms a bow shock that accelerates electrons which produces non-thermal radiation across a broad wavelength range. We discuss three wind mass loss rates: $\dot{M}_{\rm w}\sim10^{-14}$, $10^{-9}$, $10^{-6}\,M_{\odot}\,\rm yr^{-1}$ corresponding to solar-type, T Tauri and massive O/B type stars, respectively. We find that the expected radio synchrotron emission from a Jupiter-like planet is detectable by the Jansky Very Large Array and the Square Kilometer Array at ~1-10 GHz out to a distance ~ 100 pc, whereas the infrared emission is detectable by the James Webb Space Telescope out to a similar distance. Inverse Compton scattering of the stellar radiation results in X-ray emission detectable by Chandra X-ray Observatory out to ~ 150 pc. Finally, we apply our model to the upper limit constraints on V380 Tau, the first star-hot Jupiter system observed in radio wavelength. Our bow shock model provides constraints on the magnetic field, the interplanetary medium and the non-thermal emission efficiency in V380 Tau.

astro-ph.EP↗

arXiv2018

Self-sustaining star formation fronts in filaments during cosmic dawn

We propose a new model for the ignition of star formation in low-mass halos by a self-sustaining shock front in cosmic filaments at high redshifts. The gaseous fuel for star formation resides in low mass halos which can not cool on their own due to their primordial composition and low virial temperatures. We show that star formation can be triggered in these filaments by a passing shock wave. The shells swept-up by the shock cool and fragment into cold clumps that form massive stars via thermal instability on a timescale shorter than the front's dynamical timescale. The shock, in turn, is self-sustained by energy injection from supernova explosions. The star formation front is analogous to a detonation wave, which drives exothermic reactions powering the shock. We find that sustained star formation would typically propel the front to a speed of $\sim 300-700\,\rm km\,s^{-1}$ during the epoch of reionization. Future observations by the $\textit{James Webb Space Telescope}$ could reveal the illuminated regions of cosmic filaments, and constrain the initial mass function of stars in them.

astro-ph.GA↗

arXiv2018

Formation and spatial distribution of hypervelocity stars in AGN outflows

We study star formation within outflows driven by active galactic nuclei (AGN) as a new source of hypervelocity stars (HVSs). Recent observations revealed active star formation inside a galactic outflow at a rate of $\sim 15\,M_{\odot}\,\rm yr^{-1}$. We verify that the shells swept up by an AGN outflow are capable of cooling and fragmentation into cold clumps embedded in a hot tenuous gas via thermal instabilities. We show that cold clumps of $\sim 10^3\,M_{\odot}$ are formed within $\sim 10^5$ yrs. As a result, stars are produced along outflow's path, endowed with the outflow speed at their formation site. These HVSs travel through the galactic halo and eventually escape into the intergalactic medium. The expected instantaneous rate of star formation inside the outflow is $\sim 4-5$ orders of magnitude greater than the average rate associated with previously proposed mechanisms for producing HVSs, such as the Hills mechanism and three-body interaction between a star and a black hole binary. We predict the spatial distribution of HVSs formed in AGN outflows for future observational probe.

astro-ph.GA↗

arXiv2016

Ultra high energy cosmic rays from non-relativistic quasar outflows

It has been suggested that non-relativistic outflows from quasars can naturally account for the missing component of the extragalactic $γ$-ray background and explain the cumulative neutrino background through pion decay in collisions between protons accelerated by the outflow shock and interstellar protons. Here we show that the same quasar outflows are capable of accelerating protons to energies of $\sim 10^{20}$ eV during the early phase of their propagation. The overall quasar population is expected to produce a cumulative ultra high energy cosmic ray flux of $\sim10^{-7}\,\rm GeV\,cm^{-2}s^{-1}sr^{-1}$ at $E_{\rm CR}\gtrsim10^{18}$ eV. The spectral shape and amplitude is consistent with recent observations for outflow parameters constrained to fit secondary $γ$-rays and neutrinos without any additional parameter tuning. This indicates that quasar outflows simultaneously account for all three messengers at their observed levels.

astro-ph.HE↗

arXiv2016

Cumulative neutrino background from quasar-driven outflows

Quasar-driven outflows naturally account for the missing component of the extragalactic $γ$-ray background through neutral pion production in interactions between protons accelerated by the forward outflow shock and interstellar protons. We study the simultaneous neutrino emission by the same protons. We adopt outflow parameters that best fit the extragalactic $γ$-ray background data and derive a cumulative neutrino background of $\sim10^{-7}\,\rm GeV\,cm^{-2}\,s^{-1}\,sr^{-1}$ at neutrino energies $E_ν\gtrsim 10$ TeV, which naturally explains the most recent IceCube data without tuning any free parameters. The link between the $γ$-ray and neutrino emission from quasar outflows can be used to constrain the high-energy physics of strong shocks at cosmological distances.

astro-ph.HE↗

arXiv2016

Quasar-driven outflows account for the missing extragalactic gamma-ray background

The origin of the extragalactic $γ$-ray background permeating throughout the Universe remains a mystery forty years after its discovery. The extrapolated population of blazars can account for only half of the background radiation at the energy range of ~ 0.1-10 GeV. Here we show that quasar-driven outflows generate relativistic protons that produce the missing component of the extragalactic $γ$-ray background and naturally match its spectral fingerprint, with a generic break above ~ 1 GeV. The associated $γ$-ray sources are too faint to be detected individually, explaining why they had not been identified so far. However, future radio observations may image their shock fronts directly. Our best fit to the Fermi-LAT observations of extragalactic $γ$-ray background spectrum provides constraints on the outflow parameters that agree with observations of these outflows and theoretical predictions.

astro-ph.HE↗

arXiv2015

Detecting Stars at the Galactic Centre via Synchrotron Emission

Stars orbiting within 1$\arcsec$ of the supermassive black hole in the Galactic Centre, Sgr A*, are notoriously difficult to detect due to obscuration by gas and dust. We show that some stars orbiting this region may be detectable via synchrotron emission. In such instances, a bow shock forms around the star and accelerates the electrons. We calculate that around the 10 GHz band (radio) and at 10$^{14}$ Hz (infrared) the luminosity of a star orbiting the black hole is comparable to the luminosity of Sgr A*. The strength of the synchrotron emission depends on a number of factors including the star's orbital velocity. Thus, the ideal time to observe the synchrotron flux is when the star is at pericenter. The star S2 will be $\sim 0.015\arcsec$ from Sgr A* in 2018, and is an excellent target to test our predictions.

astro-ph.GA↗

arXiv2015

Probing the gaseous halo of galaxies through non-thermal emission from AGN-driven outflows

Feedback from outflows driven by active galactic nuclei (AGN) can affect the distribution and properties of the gaseous halos of galaxies. We study the hydrodynamics and non-thermal emission from the forward outflow shock produced by an AGN-driven outflow. We consider a few possible profiles for the halo gas density, self-consistently constrained by the halo mass, redshift and the disk baryonic concentration of the galaxy. We show that the outflow velocity levels off at $\sim 10^3\,\rm km\, s^{-1}$ within the scale of the galaxy disk. Typically, the outflow can reach the virial radius around the time when the AGN shuts off. We show that the outflows are energy-driven, consistently with observations and recent theoretical findings. The outflow shock lights up the halos of massive galaxies across a broad wavelength range. For Milky Way (MW) mass halos, radio observations by The Jansky Very Large Array (JVLA) and The Square Kilometer Array (SKA) and infrared/optical observations by The James Webb Space Telescope (JWST) and Hubble Space Telescope (HST) can detect the emission signal of angular size $\sim 8"$ from galaxies out to redshift $z\sim5$. Millimeter observations by The Atacama Large Millimeter/submillimeter Array (ALMA) are sensitive to non-thermal emission of angular size $\sim 18"$ from galaxies at redshift $z\lesssim1$, while X-ray observations by Chandra, XMM-Newton and The Advanced Telescope for High Energy Astrophysics (ATHENA) is limited to local galaxies ($z\lesssim 0.1$) with an emission angular size of $\sim2'$. Overall, the extended non-thermal emission provides a new way of probing the gaseous halos of galaxies at high redshifts.

astro-ph.GA↗

arXiv2014

New insights from deep VLA data on the potentially recoiling black hole CID-42 in the COSMOS field

We present deep 3 GHz VLA observations of the potentially recoiling black hole CID-42 in the COSMOS field. This galaxy shows two optical nuclei in the HST/ACS image and a large velocity offset of ~ 1300 km/s between the broad and narrow H beta emission line although the spectrum is not spacially resolved (Civano et al. 2010). The new 3 GHz VLA data has a bandwidth of 2 GHz and to correctly interpret the flux densities imaging was done with two different methods: multi-scale multi-frequency synthesis and spectral windows stacking. The final resolutions and sensitivities of these maps are 0.7" with rms = 4.6 muJy/beam and 0.9" with rms = 4.8 muJy/beam respectively. With a 7 sigma detection we find that the entire observed 3 GHz radio emission can be associated with the South-Eastern component of CID-42, coincident with the detected X-ray emission. We use our 3 GHz data combined with other radio data from the literature ranging from 320 MHz to 9 GHz, which include the VLA, VLBA and GMRT data, to construct a radio synchrotron spectrum of CID-42. The radio spectrum suggests a type I unobscured radio-quiet flat-spectrum AGN in the South-Eastern component which may be surrounded by a more extended region of old synchrotron electron population or shocks generated by the outflow from the supermassive black hole. Our data are consistent with the recoiling black hole picture but cannot rule out the presence of an obscured and radio-quiet SMBH in the North-Western component.

astro-ph.GA↗

arXiv2014

Detecting floating black holes as they traverse the gas disk of the Milky Way

A population of intermediate-mass black holes (BHs) is predicted to be freely floating in the Milky Way (MW) halo, due to gravitational wave recoil, ejection from triple BH systems, or tidal stripping in the dwarf galaxies that merged to make the MW. As these BHs traverse the gaseous MW disk, a bow shock forms, producing detectable radio and mm/sub-mm synchrotron emission from accelerated electrons. We calculate the synchrotron flux to be $\sim \rm 0.01-10\, mJy$ at GHz frequency, detectable by Jansky Very Large Array, and $\sim 10-100\,μ\rm Jy$ at $\sim10^{10}-10^{12} \,\rm Hz$ frequencies, detectable by Atacama Large Millimeter/sub-millimter Array. The discovery of the floating BH population will provide insights on the formation and merger history of the MW as well as on the evolution of massive BHs in the early Universe.

astro-ph.GA↗